Cyberpediaby Brain

Diffusion models

How an AI starts from random noise to produce an image, and what an employee can check before sharing a picture or uploading a file.

On this page
  1. 1In 30 seconds
  2. 2What it is
  3. 3Straight answers
  4. 4How it works
  5. 5What you do
  6. 6Pitfalls
  7. 7Real cases
  8. 8Sources
  9. 9Cite this article

In 30 seconds

Where it happens
In an image or video received by message, on a social network or in a document, and in public image-generation tools where files can be uploaded.
What is at stake
NIST lists among the risks harm to the integrity of information, fraud (including identity theft) and the deception of people with synthetic voices, images or videos[3].
The reflex
Before sharing an image or a video, check where it comes from; never upload confidential images or colleagues' faces to a public tool.
If it's already done
Tell the company's security team without delay, stating the tool used and what was uploaded.

What it is

A diffusion model makes an image by starting from random noise, which it refines step by step until it gets an image that matches the prompt[3]. To learn how, it follows the opposite path: images are damaged by adding noise, then the model learns to restore them. The original idea, published in 2015, is to systematically and slowly destroy the structure of data through a diffusion process, then to learn the reverse process that restores it[1]. In 2020, researchers presented high-quality image synthesis results with this class of models[2].

For an employee, the stakes are practical. These models are used to produce realistic synthetic images, which can arrive by message or be made in a public tool. NIST, the US standards institute, notes that generative AI can produce realistic images, text, audio and video, which raises new risks for trust in digital information[3].

Straight answers

What is a diffusion model?

It is a generative model that produces data, here images, by reversing a progressive scrambling. The principle draws on nonequilibrium statistical physics: the structure of data is destroyed by repeated diffusion, then the model learns to rebuild it[1]. DALL-E 2, Imagen and Stable Diffusion are image diffusion models[9].

How does a model learn to clean up noise?

It learns by doing the opposite of what it will do later. Noise is added to training images, in small amounts, until only noise is left. The model learns the reverse process, which restores the structure[1]. The authors of a 2020 paper sum up the idea this way: "Creating noise from data is easy; creating data from noise is generative modeling."[4] Once trained, the model starts from random noise and refines it little by little[3].

This cleaning takes many steps. A variant called DDIM produces high-quality samples 10 to 50 times faster, in wall-clock time, than the 2020 method, with the same training procedure[5].

How does a text prompt guide the image?

The text acts as an instruction during the cleaning. Latent diffusion models add so-called cross-attention layers to the model, which let it be guided by inputs such as text; they work in the latent space of pretrained autoencoders, which reduces computing needs compared with models that operate directly on pixels[6]. The method known as classifier-free guidance trains a conditioned model and an unconditioned model together, then combines their estimates to trade off sample quality against diversity[7]. The Imagen model relies on large general-purpose language models to encode the text, and its authors find that making them larger improves image fidelity and image-text alignment more than making the image model larger[8].

Can you recognize a generated image by eye?

Not reliably. Perceptible cues (reflections or shadows that do not match the lighting, blending artifacts) mostly serve detectors[3]. It also reports a 2024 study, covering audio, visual and audiovisual modalities, that finds performance close to chance[3]. The authors of Imagen speak of an "unprecedented degree of photorealism"[8]. NIST further writes that the effectiveness of many content transparency approaches has not yet been fully examined[3]. A reflex based on a quick look alone is therefore fragile; checking the source is less so.

Can an image carry a mark of origin?

Sometimes. NIST describes two ways to record the origin of a piece of content: digital watermarking, which embeds the information in the content itself, and metadata recording, signed or not[3]. A discreet watermark is readable by a machine and hard for a human to spot; a detector is needed, and it has a non-zero probability of false positives and false negatives[3]. The absence of a mark therefore proves nothing.

Can a model bring back an image from its training?

Yes. Researchers have shown that diffusion models memorize images from their training data and reproduce them at generation time[9].

How it works

The chain has two stages. In training, the model learns to restore images that have been scrambled[1]. In use, it starts from random noise and refines it step by step[3]; the text provided steers the result[6][7].

A more general mathematical formulation describes the scrambling as an equation that slowly turns a complex data distribution into a known one by injecting noise, and a reverse equation that removes this noise[4]. This framework encompasses earlier approaches to score-based generation and probabilistic diffusion[4]. The non-specialist reader does not need the equations: remembering that the image is born from noise removed in steps is enough.

NIST describes the journey of synthetic content in three stages: creation, publication, then consumption, meaning the moment when people receive, interpret and act on the content[3]. The employee takes part at both ends: they can create, by using a tool, and consume, by receiving an image.

What you do

  • Before sharing a surprising image or video, look for where it comes from: who published it first, in what context, and whether the sender can confirm it through another channel.
  • If the content carries a provenance mark or metadata, check them without taking them for complete proof: NIST writes that none of these techniques offers a complete solution on its own[3].
  • Faced with an urgent request backed by a voice or a video (transfer, access), check by a means other than the one used for the request. NIST cites voice cloning among the ways to get someone to facilitate a fraudulent transaction[3].
  • Never upload confidential images (plans, photographed documents, screens) to a public image-generation tool.
  • Never upload colleagues' faces. Researchers have extracted photographs of people present in their training data from diffusion models[9].
  • If in doubt about content received, pass it to the security team rather than circulating it.

Pitfalls

  • Hyperrealism. The authors of Imagen claim an unprecedented degree of photorealism, and human raters judge its samples to be on par with the COCO data for image-text alignment[8]. A convincing image is not an authentic image.
  • False confidence. NIST warns that transparency can create a false sense of trust, for example when content looks legitimate by technical measures but has been taken out of context[3].
  • Detection tools. They are not yet everywhere: NIST writes that most of these approaches may be years away from wide deployment on mobile devices[3].
  • Doctored content. NIST writes that synthetic images, voices or videos can be used to fool biometric authentication systems or people[3].
  • Memorization. The researchers conclude that diffusion models are much less private than earlier generative models such as GANs[9]. What the training set contains can therefore come back out.

Real cases

January 2023: training images extracted from state-of-the-art models

On 30 January 2023, nine researchers published a study on extracting training data from diffusion models. With a method that generates images and then filters those that match training data, they extracted over a thousand of them from state-of-the-art models, photographs of people as well as logos of trademarked companies[9]. The authors add that mitigating these vulnerabilities may require new advances in privacy-preserving training[9].

Sources

  1. arXiv, “Deep Unsupervised Learning using Nonequilibrium Thermodynamics”, 12 March 2015. https://arxiv.org/abs/1503.03585 (accessed 10 November 2026)
  2. arXiv, “Denoising Diffusion Probabilistic Models”, 19 June 2020. https://arxiv.org/abs/2006.11239 (accessed 10 November 2026)
  3. NIST, “Reducing Risks Posed by Synthetic Content: An Overview of Technical Approaches to Digital Content Transparency”, 18 November 2024. https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-4.pdf (accessed 10 November 2026)
  4. arXiv, “Score-Based Generative Modeling through Stochastic Differential Equations”, 26 November 2020. https://arxiv.org/abs/2011.13456 (accessed 10 November 2026)
  5. arXiv, “Denoising Diffusion Implicit Models”, 6 October 2020. https://arxiv.org/abs/2010.02502 (accessed 10 November 2026)
  6. arXiv, “High-Resolution Image Synthesis with Latent Diffusion Models”, 20 December 2021. https://arxiv.org/abs/2112.10752 (accessed 10 November 2026)
  7. arXiv, “Classifier-Free Diffusion Guidance”, 26 July 2022. https://arxiv.org/abs/2207.12598 (accessed 10 November 2026)
  8. arXiv, “Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding”, 23 May 2022. https://arxiv.org/abs/2205.11487 (accessed 10 November 2026)
  9. arXiv, “Extracting Training Data from Diffusion Models”, 30 January 2023. https://arxiv.org/abs/2301.13188 (accessed 10 November 2026)